gregyoung14/openmarket-btc-polymarket

OpenMarket BTC Polymarket

OpenMarket BTC Polymarket is a high-frequency research dataset pairing Binance BTC/USDT market data with Polymarket BTC binary-market order book events.

The dataset is released to support reproducible prediction-market research, feature engineering, market microstructure analysis, and backtesting.

Paper

This dataset accompanies OpenMarket: A Synchronized Polymarket-Binance Dataset for High-Frequency Prediction-Market Research (arXiv:2607.26245).

How to Cite

If you use OpenMarket, cite the paper and report the dataset version and Hugging Face revision used in your experiment.

@misc{young2026openmarketsynchronizedpolymarketbinancedataset,
  title         = {OpenMarket: A Synchronized Polymarket-Binance Dataset for High-Frequency Prediction-Market Research},
  author        = {Gregory Young},
  year          = {2026},
  eprint        = {2607.26245},
  archivePrefix = {arXiv},
  primaryClass  = {q-fin.TR},
  url           = {https://arxiv.org/abs/2607.26245}
}

Dataset-specific citation:

@misc{openmarket_dataset_2026,
  title        = {{OpenMarket BTC Polymarket}: Synchronized High-Frequency Corpus},
  author       = {Young, Gregory},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/gregyoung14/openmarket-btc-polymarket}},
  note         = {Version v0.4.3-unified; source tag v0.5.2}
}

Full entries: CITATION.md.

Repository Layout

binance_trades.parquet         # v0.1-sample (flat at repo root for Data Studio)
binance_ticks_ms.parquet
polymarket_ticks_ms.parquet
lag_pairs_ms.parquet
binance_candles_{1s,5s,1m,5m,15m,1h}.parquet
market_meta.parquet
crossover_alerts.parquet
full/                          # complete 202-snapshot archive (v0.2-full)
  <table>/date=YYYY-MM-DD/*.parquet
  metadata/<snapshot>.export_report.json
unified/                       # deduped research timeline (v0.4.3-unified)
  <table>/date=YYYY-MM-DD/*.parquet
  metadata/merge_quality_report.json
features/                      # optional ML exports (v0.4-features; sample demo on HF)
  step2_100ms/
  step2_1s/
  step3_binary_calibration/
metadata/
  snapshot_manifest.json       # full archive inventory (CDN URLs redacted)
  snapshot_manifest.tsv
README.md

Each split uses Hive-style date=YYYY-MM-DD partitions under table directories. Parquet files also include a date column (UTC, YYYY-MM-DD) for downstream filtering.

OpenMarket is in archival shutdown. No new snapshots will be collected; this repository is a fixed public research record frozen at source tag v0.5.2.

Published Splits

The operator archive contains 202 SQLite snapshot files from 2026-03-14T19:32:15Z through 2026-07-01T02:56:54Z. The inventory is in metadata/snapshot_manifest.{json,tsv} (CDN hostname redacted to cdn.example.com).

Split	Version	Description
unified/	v0.4.3-unified	Recommended. Deduped timeline (~727M rows, 504 parquet files)
full/	v0.2-full	Complete 202-snapshot per-export archive (3,312 parquet files)
(repo root)	v0.1-sample	Tiny demo split for CI and quickstarts (12 flat parquet, 9,352 rows)
features/	v0.4-features	Optional. Step2/step3 demo (one snapshot on HF); full features reproducible from unified/

unified/ backfill (v0.4.3-unified)

Adds five date partitions missing from v0.4.2-unified, sourced from staging SQLite via unified-backfill sqlite-fill:

• binance_trades/date=2026-03-23 (~191k rows)
• binance_trades/date=2026-05-15
• binance_ticks_ms/date=2026-05-15
• polymarket_ticks_ms/date=2026-05-15
• lag_pairs_ms/date=2026-05-15

Provenance: unified/metadata/sqlite_fill.json. April 22–May 12 remains a collection gap with no recoverable archive data.

Features split (v0.4-features) — optional

A full-archive features/ upload is not required for the public research record. The recommended path is to export step2/step3 features from unified/ Parquet using the published Rust and Python tooling (same pipeline that produced the v0.2.1/ model). Hugging Face currently hosts a one-snapshot demo (2 parquet files) for schema reference.

Reproduce from unified/:

cargo build -p step3-parquet-export -p binary-outcome-trainer --release
./target/release/export_step3_from_parquet \
  --parquet-root data/hf_release/unified_parquet \
  --out-dir data/hf_release/features_exports

Per-snapshot step2/step3 via SQLite (legacy path):

.venv/bin/python scripts/datasets/export_ml_features.py \
  --snapshot polymarket_btc_data_2026-03-14_193215.db.gz

See scripts/ml/README.md in the source repo.

Load the unified split:

from huggingface_hub import snapshot_download
root = snapshot_download(
    "gregyoung14/openmarket-btc-polymarket",
    repo_type="dataset",
    allow_patterns=["unified/**", "metadata/**", "README.md"],
)

Sample split (v0.1-sample)

Published as 12 flat *.parquet files at the repository root (not under sample/) so the Hugging Face Data Studio viewer can index them without traversing a split subdirectory. Download with allow_patterns=["*.parquet", "metadata/**"] or use datasets/download.py --split sample.

Source snapshot: polymarket_btc_data_2026-05-14_145928.db.gz (SQLite integrity check: ok).

Table	Rows
binance_trades	4,249
binance_ticks_ms	1,044
polymarket_ticks_ms	2,000
lag_pairs_ms	1,769
binance_candles_1s	9
binance_candles_5s	1
binance_candles_1m	104
binance_candles_5m	23
binance_candles_15m	5
binance_candles_1h	2
market_meta	146
crossover_alerts	0
Total	9,352

Schema

Each parquet file is a flat table; column dtypes are inferred from the file and surfaced by huggingface_hub.list_repo_files / pyarrow.parquet.read_schema. A reference schema follows:

binance_trades

column	dtype	description
trade_id	int64	Binance trade ID
trade_time	int64	exchange trade timestamp (ms since epoch, UTC)
price	float64	trade price (USDT per BTC)
quantity	float64	trade quantity (BTC)
quote_volume	float64	trade quantity * price (USDT)
is_buyer_maker	bool	true if buyer was the maker side
received_at	int64	collector ingest timestamp (ms since epoch, UTC)
date	string	UTC date partition derived from trade_time

binance_ticks_ms

column	dtype	description
id	int64	synthetic ID
source_ts_ms	int64	source event timestamp (ms since epoch, UTC)
ingest_ts_ms	int64	collector ingest timestamp (ms since epoch, UTC)
market_slug	string	related Polymarket market slug
price	float64	Binance mid price
best_bid	float64	Binance top bid
best_ask	float64	Binance top ask
date	string	UTC date partition derived from source_ts_ms

polymarket_ticks_ms

column	dtype	description
id	int64	synthetic ID
source_ts_ms	int64	source event timestamp (ms since epoch, UTC)
ingest_ts_ms	int64	collector ingest timestamp (ms since epoch, UTC)
market_slug	string	Polymarket market slug
asset_id	string	conditional token ID
side_label	string	"up" or "down"
event_type	string	Polymarket event type
price	float64	mid price
best_bid	float64	top bid
best_ask	float64	top ask
size	float64	depth at top of book
paired	bool	true if this tick was paired with a Binance tick
date	string	UTC date partition derived from source_ts_ms

lag_pairs_ms

column	dtype	description
id	int64	synthetic ID
paired_at_ms	int64	pairing timestamp (ms since epoch, UTC)
market_slug	string	Polymarket market slug
side_label	string	"up" or "down"
binance_tick_id	int64	FK → binance_ticks_ms.id
polymarket_tick_id	int64	FK → polymarket_ticks_ms.id
binance_source_ts_ms	int64	paired Binance source timestamp
polymarket_source_ts_ms	int64	paired Polymarket source timestamp
lead_lag_ms	int64	polymarket_source_ts_ms - binance_source_ts_ms
binance_price	float64	Binance mid at pairing time
polymarket_bid	float64	Polymarket bid at pairing time
price_delta_bps	float64	price differential in basis points
quality_flag	int64	pairing quality flag (0 = good)
date	string	UTC date partition derived from paired_at_ms

binance_candles_{1s,5s,1m,5m,15m,1h}

column	dtype	description
candle_start	int64	candle start timestamp (ms since epoch, UTC)
open	float64	open price
high	float64	high price
low	float64	low price
close	float64	close price
volume	float64	base asset volume (BTC)
quote_volume	float64	quote asset volume (USDT)
trades	int64	number of trades in the candle
date	string	UTC date partition derived from candle_start

market_meta

column	dtype	description
slug	string	Polymarket market slug
condition_id	string	on-chain condition ID
question	string	market question
end_date_iso	string	market close time (ISO 8601)
up_token_id	string	conditional token ID for "up" outcome
down_token_id	string	conditional token ID for "down" outcome
resolved_outcome	string	winning outcome, if resolved
closed	bool	true if market is closed
date	string	always "unpartitioned"

crossover_alerts

Empty in this snapshot; reserved for the strategy alert table.

Timestamp Semantics

• source_ts_ms: timestamp from the exchange or source event (ms since epoch, UTC)
• ingest_ts_ms: timestamp observed by the collector (ms since epoch, UTC)
• lead_lag_ms = polymarket_source_ts_ms - binance_source_ts_ms

Positive lead_lag_ms values indicate that the Polymarket event timestamp follows the Binance event timestamp.

Limitations

• WebSocket reconnects can create gaps.
• Collector host clocks can drift.
• Raw JSON columns are excluded from the default Parquet export.
• Top-of-book backtests may overstate executable fill quality.
• The initial sample is intentionally small and is not representative of full historical coverage.
• This snapshot is a point-in-time export; re-running the export pipeline on a later snapshot may produce slightly different row counts if the underlying recorder continued collecting between exports.

Release artifacts

Source tag:      v0.5.2
Source repo:     github.com/gregyoung14/openmarket
Dataset:         huggingface.co/datasets/gregyoung14/openmarket-btc-polymarket
Dataset version: v0.4.3-unified
Models:          huggingface.co/gregyoung14/openmarket-models
Model version:   v0.2.1 (recommended; walk-forward logistic on unified step3)
                 v0.2 / v0.1 (historical)